--- title: "Kimi K3.1 Pro: close to Claude at coding, open weights by Nov 4" description: "Moonshot released Kimi K3.1 Pro with coding scores close to Claude Opus 5.5 and promised open weights by Nov 4. Until then it runs only on Kimi's own apps and API." url: https://ihsanalapsi.dev/articles/kimi-k3-1-pro-open-weights lang: en date: 2026-10-24 updated: 2026-10-24 tags: ["AI", "Open models"] author: Ihsan Alapsi (إحسان العبسي) --- # Kimi K3.1 Pro: close to Claude at coding, open weights by Nov 4 On October 23, the Chinese company Moonshot announced Kimi K3.1 Pro, the first model in its K3.1 series, and called it its "most capable 3T-class open-weight model to date." Before you get excited, note that the weights are not out yet and the license is still unknown. ## What exactly is Kimi K3.1 Pro? Kimi K3.1 Pro is Kimi K3 after a round of post-training with much more compute than before. It has a 1-million-token context window and is available now on kimi.ai, Kimi Work, Kimi Code, and the Kimi API under these model ids: ``` kimi-k3.1-pro kimi-k3.1-pro-highspeed ``` The second one, Highspeed, is 2 to 3 times faster than K3.1 Pro according to Moonshot, and has its own model id on the API. Speed matters for coding agents because the agent waits for the model's reply at every step, so a strong but slow model holds up the whole task. Moonshot promised the full open weights by November 4, 2026. So for now you can only use the model through Kimi's own services, and there are no weights to run on your own server or fine-tune. The launch post does not mention a license at all, so do not build a plan on the assumption that it will allow everything. ## What numbers did Moonshot publish? Moonshot's numbers put Kimi K3.1 Pro less than a point behind Claude Opus 5.5 on Terminal-Bench 4.0, and 7.5 points ahead of GPT-6 Astra. They are all the company's own launch figures, some of them from its internal benchmarks, so test the model on your own tasks before you judge. | Benchmark | Kimi K3.1 Pro | For comparison | | ------------------------------------------------------ | ------------- | ------------------------------------------ | | Terminal-Bench 4.0 | 65.7% | Claude Opus 5.5: 66.4%, GPT-6 Astra: 58.2% | | Kimi Code Bench 3.0 (long-horizon, internal) | 61.2% | Kimi K3: 55.95% | | BenchCAD (with tools) | 92.7% | Claude Opus 5.5: 96.2%, GPT-6 Astra: 95.9% | | Hallucination rate on Kimi Hallu Bench Hard (internal) | 8.3% | Kimi K3: 23.4% | If these numbers hold up on real tasks, there is now an open-weight model close to the strongest closed models at coding. On BenchCAD, which asks the model to rebuild a 3D object from its images by writing CAD code, it still trails Claude Opus 5.5 by 3.5 points and GPT-6 Astra by 3.2. The hallucination rate fell from 23.4% to 8.3% on Kimi Hallu Bench Hard, Moonshot's internal test of factual accuracy. Moonshot itself says the test is built to catch models out, so the number is not the error rate you will see in daily use. The direction still matters for coding agents, because an agent that hallucinates invents functions and files that do not exist, and you spend time cleaning up after it instead of saving time. According to the announcement, the model also works on full game projects in Unity, Unreal Engine, and Godot. For parametric 3D modeling, it writes and runs scripts for Blender and CAD software without MCP connectors. The announcement also covers complete finance workflows. ## Why did they remove cyber data from training? Moonshot removed the cyber data from Kimi K3.1 Pro's training because, in its own words, it aims "not to release open weights that would extend the absolute frontier of cyber capability." So it removed cyber-related data from the mid-training corpus and from the reinforcement learning (RL) environments for the open-weight release, and kept the general coding tasks. On xdev-bench-hard, Moonshot's internal test of finding and exploiting vulnerabilities, the checkpoint without these measures solved 90.5% of the tasks, and the final K3.1 Pro solved 81.0%. The company accepted a 9.5-point loss in this area for the decision. You can limit a closed model's capabilities through the API, but open weights spread and you cannot take them back, so the company reduced the model's offensive capability before releasing it. The announcement does not mention a stronger cyber version on the API: the 81.0% belongs to the final K3.1 Pro, and the 90.5% to an internal training checkpoint. If you work in security or build defensive tools, do not expect top performance in vulnerability research. Moonshot itself says the model's cyber capability is still behind the strongest closed models that have been publicly available for at least three months. ## What does this mean for you if you build coding agents? If you build coding agents, Kimi K3.1 Pro is worth testing on the API now, but any plan to host it yourself has to wait for the weights and the license. First, test the model on the API and measure it on your own tasks, not on the launch table. If you build a coding agent, try `kimi-k3.1-pro-highspeed` for tasks that need speed, and compare the results with the model you use now. Second, cost. Open models usually come out cheaper than closed ones once their weights are out, because you can run them on your own infrastructure. If these coding numbers hold up, there is a serious alternative to Claude for coding tasks, and that changes the cost math for anyone running coding agents in large volumes. Third, do not rush. Until the weights land (due by November 4) along with the license terms, everything is talk. Earlier open models have shipped with restrictive licenses that did not allow free commercial use, so do not build a production plan on assumptions. ## Where does this fit among open-model news? Kimi K3.1 Pro is one of several open models released this month that compete on coding. Moonshot compares it directly with Claude Opus 5.5 and GPT-6 Astra, and Claude Opus 5.5 is still ahead even in Moonshot's own table. We recently looked at [Jais 3](/articles/jais-3-arabic-open-models), which focuses on Arabic. Its numbers also come from its own developers, like Kimi's numbers now. We also looked at [the Muse Spark Open license](/articles/meta-muse-spark-open-licence) from Meta, which allows commercial use but with conditions. That is why the license matters as much as the weights. Models are also leaving service, like [the models removed from Codex and Copilot](/articles/ai-model-retirements-codex-copilot) this month, which is one more reason to test more than one model before you tie your work to one. ## How do you try it on the API before the weights land? You try Kimi K3.1 Pro on the Kimi API with the same tasks you give your current model, and compare the results. In order: 1. Open the [official announcement](https://www.kimi.ai/blog/kimi-k3-1-pro) and check the details of both models and their model ids. 2. Try `kimi-k3.1-pro` on a real coding task from your work, not a ready-made example: take a ticket from your projects with a lot of context and see how it handles it. 3. Try `kimi-k3.1-pro-highspeed` on interactive tasks that need a quick reply, such as completion while you type or quick agent questions, and compare speed and quality. 4. Measure it yourself: same task, same prompt, on your current model and on K3.1 Pro, and note which one produced working code on the first try. About the 1-million-token context: quality usually drops as more of the context is filled, so do not assume the model remembers everything across a million tokens with the same accuracy. Try it on a large repository and see whether it uses the long context or gets lost in it. Before you think about switching your agent's model, answer two questions. Does the model follow complex instructions the way you want, or does it cut corners? And does the lower hallucination rate Moonshot announced show up on your tasks, meaning fewer invented functions and files than your current model? ## What should you do now? Three steps: 1. Try the model on the API now from the [official announcement](https://www.kimi.ai/blog/kimi-k3-1-pro), and measure it on your own tasks. 2. Wait for the weights (due by November 4), and read the license terms as soon as they are out, before you plan any self-hosting. 3. If you build coding agents, compare cost and performance with your current model, because this news may change your math. What do you think: if the weights come out under a good license, would you move your coding agent to an open model, or stay with a closed one? ## FAQ **When do the open weights come out?** By November 4, 2026, according to Moonshot. Until then the model is only available on Kimi's own apps and API, and there is nothing to download. **Can I self-host it right now?** No. The weights are not out yet, and the launch post does not mention a license at all. Do not plan on self-hosting until both are published. **How does it compare to Claude on coding?** Less than a point behind Claude Opus 5.5, by Moonshot's own numbers: 65.7% on Terminal-Bench 4.0 versus 66.4%, and ahead of GPT-6 Astra at 58.2%. These are the vendor's launch figures, so test it on your own tasks before you judge. **How much did removing the cyber data cost the model?** 9.5 points on xdev-bench-hard, Moonshot's internal test of finding and exploiting vulnerabilities: the checkpoint without the cleanup scored 90.5%, and the final K3.1 Pro scored 81.0%.